Research On Mushroom Image Classification Based On ConvNeXt

In daily life,mushrooms are highly valued as a cooking ingredient due to their high nutritional value,and they are greatly loved by the public.However,wild mushrooms have gradually entered people’s field of vision and have even been served on dining tables.To prevent people from mistakenly consuming toxic mushrooms,it is crucial to establish an effective mushroom classification and identification system.Currently,traditional manual mushroom classification methods suffer from drawbacks such as low efficiency,high cost,and specialized expertise.In recent years,Convolutional Neural Networks(CNN)have made rapid advancements in the field of image classification,with the ConvNeXt model showing excellent performance in image classification tasks.Therefore,applying ConvNeXt to mushroom image classification holds significant importance in overcoming the limitations of manual methods and continuously improving the accuracy of mushroom image classification.Additionally,in order to deploy the mushroom image classification model in mobile services and industrial production,designing a lightweight model is essential.The main research objectives of this paper are as follows:(1)In response to the insufficient variety and imbalanced data in current research on mushroom image classification,this paper employs a supervised data augmentation method.Multiple variations,including geometric and color transformations,are applied to the original mushroom images in the dataset,resulting in the creation of the MO106＿2 mushroom image dataset.Experimental verification demonstrates that the model trained on the MO106＿2 dataset achieves a 1.7% higher accuracy compared to the model trained on the original dataset.(2)Addressing the drawbacks and limitations of traditional manual mushroom classification methods,as well as the issue of low accuracy in multi-class mushroom image classification research,this paper applies the ConvNeXt model to the study of mushroom image classification.Utilizing transfer learning methods,experimental results demonstrate that the ConvNeXt-T model achieves a classification accuracy of90.5%.Subsequently,through two sets of comparative experiments,the advantages of transfer learning methods in terms of model training efficiency and the higher classification accuracy of the ConvNeXt model compared to other models are validated.(3)In response to the low accuracy issue of lightweight models in multi-class mushroom image classification research,a lightweight Conv X-Shuffle Net model is proposed by combining ShuffleNetV2 with ConvNeXt.This study optimizes the structure and algorithms of the ShuffleNetV2 model and presents three schemes:Conv X-Shuffle Net-1,Conv X-Shuffle Net-2,and Conv X-Shuffle Net-3.Experimental data demonstrates that their classification accuracies are 82.0%,82.6%,and 81.9%,respectively,all surpassing other models.Furthermore,these three schemes exhibit different strengths and weaknesses in terms of accuracy,FLOPs,and params.